Is it possible to use two-dimensional positional encoding in transformers working with images?

Yes, two-dimensional positional encoding can and often should be used in transformers working with images. Unlike textual data, where positional encoding is usually one-dimensional (a sequence of tokens), images have a two-dimensional structure (width and height).

Two-dimensional positional encoding allows the model to consider the spatial coordinates of pixels or patches, which improves the understanding of the image structure.

An example approach:

• Create separate positional vectors for the X and Y axes.
• Add or concatenate these vectors to obtain the final positional encoding for each patch.

This helps the transformer to consider the arrangement of elements in two-dimensional space, which is critical for computer vision tasks.